A METHOD AND SYSTEM FOR PREDICTING A VEHICLE SPEED

Abstract

This invention relates to a method and system for Vehicle Speed Prediction. It attempts to solve mileage problems in autonomous electric vehicles with objective of predicting vehicle speed accurately. Accurate vehicle speed prediction plays crucial role in transportation industry. The system(100) is having data processor(101) in communication with hybrid Embedded GRU architecture. Said hybrid Embedded GRU architecture includes plurality of convolution layers(102A,102B), plurality of GRU layers(103A,103B) and one dense layer(104) resulting vehicle speed prediction(105). The data processor (101) is fed with at least four wheel speed data(101A) of vehicle with breaking and non-breaking conditions on different types of road including snow, high-mountain, dry road, straight dry asphalt, ice and curve circles for processing the same, which is subsequently fed to convolution layers(102A,102B), embedded GRU layers(103A,103B) and dense layer(104) sequentially to obtain desired output(105). The invention is energy efficient and highly accurate for vehicle speed prediction causing increase in vehicle mileage. (Figure 1 & 3)

Specification

Description:"A METHOD AND SYSTEM FOR PREDICTING A VEHICLE SPEED"

FIELD OF INVENTION:

[001] This invention relates to a method and system for predicting a vehicle speed. This system is having Energy efficient hybrid Embedded GRU architecture for the speed prediction.

BACKGROUND/PRIOR ART OF THE INVENTION

[002] Background description includes information that may be useful in understanding the present invention. It is not an admission that any of the information provided herein is prior art or relevant to the presently claimed invention, or that any publication specifically or implicitly referenced is prior art.

[003] Vehicle speed prediction is an essential component of modern automotive technology, especially in developing intelligent transportation systems (ITS) and self-driving cars. Accurate speed prediction is critical for increasing safety and improving fuel efficiency. For example, in autonomous vehicles, predicting speed aids in making real-time decisions, such as adjusting acceleration or braking, to maintain safe distances from other vehicles and navigate complex driving situations. This results in smoother driving experiences, less congestion, and lower emissions.

[004] Advanced deep learning algorithms, such as LSTMs, GRUs, and Transformers, are increasingly being used to improve the accuracy of speed predictions by analysing massive amounts of data. In addition to improving safety and efficiency, vehicle speed prediction is important in energy management systems, particularly in electric vehicles (EVs). These systems can extend the vehicle's range by optimizing power consumption based on speed changes. Therefore, the electric vehicles' mileage remains a significant concern. Hence, there is a need for compact and efficient deep neural network system for electric vehicular technologies that can run on onboard sensor computing edge devices such as MCUs or IoTs.

[005] The present invention attempts to reduce a battery's energy consumption, wherein increasing a vehicle's mileage is key. So, the vehicle speed prediction has been selected as an application and an energy-efficient and high-speed hybrid embedded GRU architecture has been built. Several studies are available for vehicle speed prediction using different deep neural networks, including LSTMs, Bi-GRU, GRU, etc. However, all these studies are based on image data, whereas the present invention uses only the wheel speeds of the vehicle as input data.

[006] Some literatures on vehicle speed prediction can be discussed hereinunder, in which very few models are deployed on various hardware platforms.

[007] Literature : Multi-range vehicle speed prediction using vehicle connectivity for enhanced energy efficiency of vehicles
Apr 8, 2021 - THE REGENTS OF THE UNIVERSITY OF MICHIGAN

[008] It talks about an integrated speed prediction framework based on historical traffic data mining and real-time V2I communications for CAVs. The present framework provides multi-horizon speed predictions with different fidelity over short and long horizons. The present multi-horizon speed prediction is integrated with an economic model predictive control (MPC) strategy for the battery thermal management (BTM) of connected and automated electric vehicles (EVs) as a case study. The simulation results over real-world urban driving cycles confirm the enhanced prediction performance of the present data mining strategy over long prediction horizons. Despite the uncertainty in long-range CAV speed predictions, the vehicle level simulation results show that 14% and 19% energy savings can be accumulated sequentially through eco-driving and BTM optimization (eco-cooling), respectively, when compared with normal-driving and conventional BTM strategy.
[009] Literature 1: "Energy-Efficient High-Speed Architecture for Vehicle Speed Prediction Using Microcontrollers." by Mohammed Rafi Shaik et al. (2023)-

[0010] This comprehensive study discusses the 4L-LSTM, 4L-GRU, and 4L-eGRU architecture for the vehicle speed prediction deployed on the AURIX TC277 and LPC 1768 MCU platform. They achieved an MAE of 0.75 for 4L-eGRU, with an inference latency of 0.523 ms with an energy consumption of 0.365 mJ on AURIX TC277 and LPC 1768, an inference latency of 0.974 ms with an energy consum ption of 0.474 mJ. In comparison, proposed model has an inference latency of 0.386 ms with an energy consumption of 0.204 mJ on Aurix TC277 MCU, and on LPC 1768, the latency was 0.530 with an energy consumption of 0.28.

[0011] Literature2: "Ego-Vehicle Speed Prediction Using a Long Short-Term Memory Based Recurrent Neural Network." by K. Yeon et al. (2019)-

[0012] The LSTM-based RNN model estimates ego-vehicle speed for 15 seconds using inputs from the previous 30 seconds. The model was tested with real driving data in three scenarios: car-following, sharp curve road, and whole path. They have evaluated the model performance using R2 and RMSE loss functions.

[0013] Literature 3: "Vehicle speed prediction using a Markov chain with speed constraints." by J. Shin et al. (2018)-

[0014] The overall structure of vehicle speed prediction utilizing a Markov chain with speed constraints is composed of three stages: data collecting, data encoding, and model combination. The vehicle utilized in the experiment is a Kia Sorrento with a 2.6 GHz Intel Core i7-6600U CPU personal laptop. They have evaluated R2 and RMSE on different routes with an R2 of 0.813 and RMSE of 3.73. In contrast, the proposed architecture R2 error is 0.998, the higher the value the best the model with an RMSE (loss function) is 1.072.

[0015] Literature 4: "A big data-based deep learning approach for vehicle speed prediction." by Cheng, Z et al. (2017)-

[0016] This research presents a big data-based deep learning system for addressing the VSP based on numerous characteristics such as driver behaviour, traffic, route type, and weather conditions while predicting vehicle speed. This BDDL-SP technique is suitable for any route type. In this article the authors have evaluated the prediction accuracy using mean absolute error (MAE), root mean squared error (RMSE) and reported MAE as 5.64, RMSE as 7.73. Here, the instant invention also performed best regarding the accuracy of the speed predictions.

[0017] Vehicle speed is influenced by a variety of factors, including driver conduct, weather conditions, traffic circumstances, route type, route curvature, and so on. The relationship between these elements and vehicle speed is extremely nonlinear and complex.

[0018] However, these traditional methods have several drawbacks as follows:
- Latency,
- Memory,
-Energy consumption.

[0019] Despite good accuracy, most of the said architectures are limited in their applicability on resource-constrained platforms like MCUs. Therefore, only four-wheel sensor data was used as input features to develop system of present invention on different types of road conditions like snow, ice, high-mu, and straight dry asphalt. The proposed architecture is best suited for vehicle speed prediction where the resources are the major concern from the hardware perspective without neglecting accuracy.
OBJECTS OF THE INVENTION

[0020] Primary object of the invention is to provide a method and system for predicting a vehicle speed.

[0021] Another object of the invention is to provide a miniaturized, accurate, energy-efficient vehicle speed prediction hybrid deep neural network system on the real-time edge computing IOT devices (MCUs) resulting in improved mileage performance / fuel economy of an autonomous electric vehicle.

[0022] One another object of the invention is to provide a method and system for predicting a vehicle speed , which obviates short comings of the prior arts.

[0023] Still another object of the invention is to provide a system for predicting a vehicle Speed, which is having simple architecture.

[0024] Yet another object of the invention is to provide a method and system for predicting a vehicle speed , which serves its purpose effectively.

[0025] These and other objects and advantages of the present invention will be apparent to those skilled in the art after a consideration of the following detailed description taken in conjunction with the accompanying drawings in which a preferred form of the present invention is illustrated.

SUMMARY OF THE INVENTION

[0026] One or more drawbacks of conventional systems and process are overcome, and additional advantages are provided through the apparatus/composition and a method as claimed in the present disclosure. Additional features and advantages are realized through the technicalities of the present disclosure. Other embodiments and aspects of the disclosure are described in detail herein and are considered to be part of the claimed disclosure.

[0027] Nowadays, the electric vehicle's mileage remains a significant concern. To enhance the battery operational time, there is need for compact and efficient deep neural networks for electric vehicular technologies, so a hybrid embedded GRU architecture with around 1444 parameters has been proposed together with the inference mean absolute error (MAE) of 0.5, and an energy usage of 0.204 mJ on the Aurix TC277 MCU.

[0028] Among the available DNNs, RNNs like Long short-term memory (LSTM) and gated recurrent units (GRUs) are highly suitable for handling temporal sequences of data, and CNN is very much suitable for special features, especially for prediction tasks. In the present invention, a DNN model has been provided comprising of two CNN layers, two Embedded GRU layers, and one fully connected layer, with a dataset of four-wheel speeds on different road conditions like snow, ice, high-mu, and straight dry asphalt for training. The Embedded GRU is a simplified version of GRU.

According to the invention, there is provided a method for predicting a vehicle speed, the method comprising steps of:
- continuously monitoring speed of a plurality of wheels of a vehicle to generate a vehicle speed data (100A);
- processing the vehicle speed data (100A) of the plurality of wheels with breaking and non-breaking conditions on different types of roads through a data processor (101) for refining the data to structured data in a tensor format thereby forming a processed data;
- feeding the processed data to a plurality of convolution (CNN) layers sequentially for filtration to disregard redundant information from the processed data and to generate an output containing filtered data constituting containing key vehicle speed characteristics;
- feeding the output of a last layer of the plurality of convolution layers to a plurality of eGRU layers in a sequence for condensing the output data thereby generating a condensed data;
- providing the condensed data from a last layer of the plurality of eGRU layers to a connected dense layer (104), the connected dense layer transforms the condensed data having multi-dimensional information into a single and continuous value thereby predicting the vehicle speed.

The processed data is fed to a first convolution (CNN) layer (102A) of the plurality of convolution layers to generate a filtered data as a first output, the first output is inputted to a subsequent convolution layer (102B) of the plurality of convolution layers for performing convolution to generate a further filtered data as a subsequent output;
wherein the subsequent output of the subsequent convolution layer (102B) is fed as input to a first eGRU layer (103A) for processing the data sequentially across time series to generate a first output of the first eGRU layer (103A) of the plurality of eGRU layers;
wherein the first output of the first eGRU layerl (103A) of the plurality of eGRU layers is applied as an input to a subsequent eGRU layer (103B) of the plurality of eGRU layers for refining temporal relationships between features and condensing the information for the prediction;
and subsequent output of the subsequent eGRU layer/cell (103B) is provided as input to the connected dense layer (104) to obtain the prediction speed of the vehicle (105).

The convolution layer constitutes a one dimensional layer.

The vehicle speed data is monitored on a variety of road surfaces, including dry road, straight dry asphalt, ice, snow, and curve circles.

The eGRU layers , each of which comprises of an update gate (Zt) and has soft sign (S1) as their activation function.

The step of processing the vehicle speed data includes identifying missing values, outliers, and discrepancies; and wherein the missing values are imputed using techniques comprising mean imputation and interpolation.

The vehicle speed data comprises of a speed data of a front left (FL), a front right (FR), a rear left (RL), and a rear right (RR) wheels as input.

The plurality of convolution (CNN) layers derive a spatial information including road textures, curves, and other relationships between the speed of the front and the rear wheels., , in which the plurality of eGRU layers capture temporal dynamics in vehicle speed data while compensating for the sequential character of the input.

The plurality of convolution (CNN) layers filter the input data and convolve it across spatial dimensions to extract local features, in which the output from the last layer of the plurality of CNN layers is passed into the eGRU layers, comprising of recurrent units that record sequential dependencies.

The eGRU layers use gating methods to regulate the input flow, allowing the system to selectively update the data at each timestep.

Further, according to this invention there is provided a system for predicting a vehicle speed (100) comprising of data processor (101) for refining data to structured data in tensor format thereby forming a processed data in communication with hybrid Embedded GRU architecture, wherein the hybrid Embedded GRU architecture includes a plurality of convolution layers (102A,102B) for filtration of the processed data to disregard redundant information from the processed data and to generate an output containing filtered data constituting key vehicle speed characteristics, a plurality of embedded GRU layers (103A,103B) for condensing filtered data and at least one dense layer (104) for transforming the condensed data having multi-dimensional information into a single and continuous value thereby predicting the vehicle speed, in which the data processor (101) is fed with at least four wheel vehicle speed data (101A) with breaking and non-breaking conditions on different types of roads.

The convolution layer constitutes one dimensional layer.

The vehicle speed data is monitored on a variety of road surfaces, including dry road, straight dry asphalt, ice, snow, and curve circles.

The eGRU layers , each of which comprises of an update gate (Zt) and has soft sign (S1) as their activation function.

The vehicle speed data includes identifying missing values, outliers, and discrepancies; and wherein the missing values are imputed using techniques comprising mean imputation and interpolation.


The vehicle speed data comprises of a speed data of a front left (FL), a front right (FR), a rear left (RL), and a rear right (RR) wheels as input.

The plurality of convolution (CNN) layers derive a spatial information including road textures, curves, and other relationships between the speed of the front and the rear wheels.

The plurality of convolution (CNN) layers filter the input data and convolve it across spatial dimensions to extract local features, in which the output from the CNN layers is passed into the eGRU layers, comprising of recurrent units that record sequential dependencies.

The plurality of eGRU layers use gating methods to regulate the input flow, allowing the system to selectively update the data at each timestep.


[0029] Various objects, features, aspects, and advantages of the inventive subject matter will become more apparent from the following detailed description of preferred embodiments, along with the accompanying drawing figures in which like numerals represent like components.

[0030] It is to be understood that the aspects and embodiments of the disclosure described above may be used in any combination with each other. Several of the aspects and embodiments may be combined to form a further embodiment of the disclosure.

[0031] The foregoing summary is illustrative only and is not intended to be in any way limiting. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features will become apparent by reference to the drawings and the following detailed description.

BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS

[0032] The illustrated embodiments of the subject matter will be best understood by reference to the drawings, wherein like parts are designated by like numerals throughout. The following description is intended only by way of example, and simply illustrates certain selected embodiments of devices, systems, and processes that are consistent with the subject matter as claimed herein, wherein:-

[0033] Figure 1 Illustrates: A system for predicting a vehicle Speed according to the present invention.

[0034] Figure 2 Illustrates: The gating mechanism and activation functions of gated recurrent unit (GRU) and embedded gated recurrent unit (eGRU) of the present invention.

[0035] Figure 3 Illustrates: Flow chart regarding method for predicting a vehicle Speed in accordance with the present invention.

[0036] Figure 4 Illustrates: Preferred embodiment of method for predicting a vehicle Speed in figure 3 according to the present invention.

[0037] The figures depict embodiments of the disclosure for purposes of illustration only. One skilled in the art will readily recognize from the following description that alternative embodiments of the structures and methods illustrated herein may be employed without departing from the principles of the disclosure described herein.

DETAILED DESCRIPTION OF THE INVENTION WITH REFERENCE TO THE ACCOMPANYING DRAWINGS

[0038] While the embodiments of the disclosure are subject to various modifications and alternative forms, specific embodiment thereof have been shown by way of example in the figures and will be described below. It should be understood, however, that it is not intended to limit the disclosure to the particular forms disclosed, but on the contrary, the disclosure is to cover all modifications, equivalents, and alternative falling within the scope of the disclosure.

[0039] Accurate vehicle speed prediction plays a crucial role in advanced driver assistance systems, Intelligent Transportation Systems, Autonomous Electric Vehicles, and so on. Nowadays, the transportation industry is rapidly moving towards battery vehicles. Most of the on-ground transportation vehicles emit greenhouse gases, as fossil fuels release carbon dioxide. This causes a decrease in air quality, impacting human lives directly. The industry is moving towards battery vehicles or hybrid electric vehicles to reduce the use of internal combustion engines, which cause air pollution. In intelligent braking systems, accurate vehicle speed prediction is crucial for boosting stability and efficiency, particularly in electric and autonomous vehicles.

[0040] Therefore, the present invention has been proposed, which attempts to solve mileage problems in autonomous electric vehicles with the prediction of vehicle speed.

[0041] By precisely predicting the vehicle's speed, modern systems can adjust braking pressure in real time to prevent wheel lock-up, skidding, or delayed braking responses. This enhances vehicle stability, shortens stopping distances, and reduces the risk of accidents, especially in emergencies or challenging conditions. Accurate speed prediction improves the reliability and responsiveness of advanced braking systems, leading to greater safety.

[0042] Vehicle speed is subjected to many factors, such as driver behaviour, weather conditions, traffic conditions, route type, route curvature, and so on. The mapping between these factors and vehicle speed is highly nonlinear and complex. However, these traditional methods have several drawbacks i.e. 1) Latency, 2) Memory, and 3) Energy consumption. Despite good accuracy, most of the said architectures are limited in their applicability on resource-constrained platforms like MCUs. Therefore, the invention uses only four-wheel sensor data as input features to develop system of invention on different types of road conditions like snow, ice, high-mu, and straight dry asphalt. The proposed architecture is best suited for vehicle speed prediction where the resources are the major concern from the hardware perspective without neglecting accuracy.

[0043] In order to design the accurate VSP system, the four-wheel speeds of the data has been used along with the collection of the dataset on different road conditions like snow, ice, high-mu, and straight dry asphalt, curved circles, and from different locations. The feature in the dataset comprises of front left, front right, rear left and rear right. The data preprocessing step includes identifying missing values, outliers, and discrepancies. Missing values are imputed using acceptable techniques such as mean imputation and interpolation. Outliers, which could have a major impact on the system's performance, are detected using statistical approaches such as the z-score. Data inconsistencies, such as contradicting or unrealistic values, are carefully investigated and resolved.

[0044] The next important segment before system training is selecting the required features from the dataset. The goal is to select a subset of features that capture the critical information while reducing the dataset's dimensionality. To accomplish this, a recursive feature elimination (RFE) technique was utilized. The recursive feature elimination (RFE) is an iterative procedure that starts with all features and then eliminates the least important ones in each iteration until the desired number of features is reached. To make the dataset to be balanced, sliding window technique can be used without restricting scope of the invention to the same so as to increase the scarcity of data for specific road types such as snow and ice.

[0045] The invention is a low complexity, high speed and energy efficient hybrid architecture for vehicle speed prediction.

[0046] The novel/inventive features of vehicle speed prediction hybrid Embedded GRU architecture design and deployment lie in several key aspects as follows:

- System design, and
- Deployment on resource-constrained edge computing devices.

[0047] It starts with data pre-processing, using the sliding window technique to boost the data size. While pre-processing, only four-wheel speed data is used as input.

[0048] The invention involves the following:

- A hybrid Embedded GRU architecture, which comprises of a plurality of CNNs, a plurality of Embedded GRUs, and one dense layer, as shown in Figure 1.
- System performance is estimated using different loss functions, namely MAE, MSE, RMSE, R2, and explained variance score (EVS).
- There is deployment of system of present invention on resource-constrained AURIX TC277 and NXP LPC1768 MCUs. Further, there is deployment of float32 and fixed32 on both the MCUs and to reduce resource consumption further, float16 architecture has been deployed on LPC 1768.
- All the systems are compared with the state-of-the-art RNN models like LSTM, GRU, and eGRU.

[0049] Reference may be made to the accompanying figures, wherein the figure 1 illustrating a system for predicting a vehicle speed (100) comprising of data processor (101) in communication with hybrid Embedded GRU architecture. Here, the hybrid Embedded GRU architecture includes a plurality of convolution layers (102A,102B), a plurality of GRU layers (103A,103B) and one dense layer (104) one after another sequentially in order to provide vehicle speed prediction (105). The data processor (101) is fed with wheel speed data (101A) of a vehicle with breaking and non-breaking conditions on different types of roads.

Thus, the system comprises of:
-Convolutional Layers: at least Two 1D convolution layers with ReLU activa-tions.
-Recurrent Layers: at least Two eGRU cells (a custom variant of GRU) that handle sequential data.
-Fully Connected Layer: A linear layer for output prediction.The present invention continuously monitors speed of a plurality of wheels of a vehicle to generate a vehicle speed data (100A).
The data processor (101) refines the data to structured data in a tensor format thereby forming a processed data.
The processed data is fed to a plurality of convolution (CNN) layers sequentially for filtration to disregard redundant information from the processed data and to generate an output containing filtered data constituting key vehicle speed characteristics.
The output of a last layer of the plurality of convolution layers is fed to a plurality of eGRU layers in a sequence for condensing the output data thereby generating a condensed data.
The condensed data is provided from a last layer of the plurality of eGRU layers to a connected dense layer (104). The connected dense layer transforms the condensed data having multi-dimensional information into a single and continuous value thereby predicting the vehicle speed.

Here, the processed data is fed to a first convolution (CNN) layer (102A) of the plurality of convolution layers to generate a filtered data as a first output, the first output is inputted to a subsequent convolution layer (102B) of the plurality of convolution layers for performing convolution to generate a further filtered data as a subsequent output.

The subsequent output of the subsequent convolution layer (102B) is fed as input to a first eGRU layer (103A) for processing the data sequentially across time series to generate a first output of the first eGRU layer (103A) of the plurality of eGRU layers.
The first output of the first eGRU layerl (103A) of the plurality of eGRU layers is applied as an input to a subsequent eGRU layer (103B) of the plurality of eGRU layers for refining temporal relationships between features and condensing the information for the prediction.
Subsequent output of the subsequent eGRU layer/cell (103B) is provided as input to the connected dense layer (104) to obtain the prediction speed of the vehicle (105).

The convolution layer forms a one dimensional layer.
In the present invention, 1D convolutional layers serve to extract important features from the time-series input, which is crucial for capturing patterns in vehicle speed over sequences.
It is particularly advantageous in time-series analysis for extracting local spatial features, reducing model complexity, and enabling parallel processing, making them a powerful tool for real-time and resource-constrained applications.
Its advantages can be discussed hereinbelow:
• 1D convolutions effectively capture short-term dependencies and local patterns within a time sequence, which is particularly helpful for recognizing patterns like spikes or trends over small intervals. This makes it ideal for tasks like vehicle speed prediction, where local fluctuations in data provide valuable insights into broader patterns;
• Compared to recurrent layers like LSTMs or GRUs, 1D convolution layers require fewer computations, as they don't have to maintain a memory state across time steps. This makes them faster to train and run, especially beneficial for microcontrollers or embedded systems or applications requiring low-latency predictions;
• 1D convolutions allow for parallel computation since they don't depend on sequential processing like RNNs. This allows for faster training and inference on modern hardware, which are optimized for parallel operations;
• By using different kernel sizes, 1D convolutions can capture patterns at various scales. Small kernels capture fine details over shorter sequences, while larger kernels capture broader trends over more extended sequences, providing multi-scale insights in the data;
• In wheel speed data from different sensors, 1D convolution can learn relationships across multiple input channels at each time step, allowing the model to capture interactions between variables effectively.
The input shape for the system is (X, 5, 15, 1).
Let X be the number of samples.
Processed Data Shape:
After the sliding window transformation, each sequence for training (input X_Train_Tensor) has the shape (num_samples, num_steps, num_features,
where:
 num_samples is the number of sliding windows or sequenc-es created (X)
 num_steps is the window size (which is 5).
 num_features is the number of input features.
 1 is for the depth (since it's 1D data).

[0050] Before passing the data through the convolutional layers for filtration, there is reshaping of the data to match the expected input format for nn.Conv1d. PyTorch's Conv1d expects input in the shape of (num_samples, num_steps, num_features).

[0051] For the first convolution layer (102A) the input shape is (X, 5, 15), Performs a 1D convolution with num_features =3 ,with an output channels as 16, and a kernel size of 1. This kernel only looks at individual time steps. The first convolution layer generates an output shape of (X,16,5).

[0052] The output of the first convolution layer (102A) is applied as input to the second convolution layer (102B) with a size of (X,16,5). The second conv1D performs convolution with input channels as 16 and generates an output of 8 with a kernel size of 7. This kernel looks at a wider window (7 time steps). The second convolution layer (102B) generates an output of (X,8,9).

[0053] The output of second convolution layer (102B) is fed as input to the first eGRU layer (103A) and the eGRU processes data sequentially across time steps. This eGRU cell comprises of input size of 8 and generates an output size (hidden_size) of 8.

[0054] The output shape of the first eGRU cell (103A) is (X,8,9) which serves as input to the second eGRU cell (103B). The second eGRU cell (103B) has a hidden size of 5, further refining the temporal relationships between features, condensing the information for the final prediction.

[0055] These eGRU cells help the system to learn how the vehicle speed evolves over time, enabling it to predict the future speed with high precision. Since eGRU is more efficient, this can be done with minimal delay, making it a suitable choice for real-time applications like braking systems.

[0056] The output of the second eGRU cell (103B) is (X,5,9) which serves as input to the fully connected layer (104) to generate a single output, which is the output speed of a vehicle constituting vehicle speed prediction (105).

[0057] Development of invention can be discussed hereinunder:

Data Pre-processing:
[0058] During the data pre-processing, the dataset was thoroughly analysed to understand its properties and dispersion. The dataset included vehicle speed recordings from a variety of road surfaces, including dry road, straight dry asphalt, ice, snow, and curve circles. The data was collected in real-time from different locations and different vehicles travelling in various road types. The first phase in data analysis involved investigating the dataset's structure and features. The mean, standard deviation, minimum, maximum, and quartiles were determined for the vehicle speed variable.

[0059] The data pre processing step includes identifying missing values, outliers, and discrepancies. Missing values, were imputed using acceptable techniques such as mean imputation and interpolation. Outliers, which could have a major impact on the system's performance, were detected using statistical approaches such as z-score. Data inconsistencies, such as contradicting or unrealistic values, were carefully investigated and resolved.

Features Selection:
[0060] The next important segment before system training is selection of the required features from the dataset i.e. to select a subset of features that capture the critical information while reducing the dataset's dimensionality. To accomplish this, an RFE (recursive feature elimination) technique was utilized and it is an iterative procedure which starts with all features and eliminates the least important ones in each iteration till the desired number of features are reached. After selecting the required features, data augmentation was employed, which is crucial for increasing the performance of the hybrid architecture, particularly for the unbalanced dataset. A data augmentation method known as the sliding window technique was employed to increase the scarcity of data for specific road types such as snow and ice. Further, only four-wheel speed data [front left (FL), front right (FR), rear left (RL), rear right (RR)] was used as input features while training the system and ground_truth as GPS value in this exemplary embodiment. The speed data is taken from at least four wheel vehicle.

[0061] The hybrid architecture has been trained for vehicle speed prediction. This architecture is trained with an appropriate optimisation technique, such as Adam, while system training, with the maintenance of a bach size of 9, with L1 Loss as a loss function, continued for 100 epochs. After training and testing, the system got a mean absolute error as 0.5. The system is developed using a Pytorch environment with a graphical processing unit (GPU) Nvidia Quadro P4000.

System Architecture:
[0062] The hybrid embedded GRU system [100] comprises of a plurality of convolution layers (CNNs) [102A,102B], a plurality of embedded GRU layers (eGRUs) [103A,103B] and one dense (FC) layer [104] as shown in figure 1. The system inputs are wheel speed data of a four wheeler with breaking and non-breaking conditions on different road types. The Embedded GRU i.e. eGRU is introduced as an optimized form of GRU as shown in figure 2. The GRU works with two gate mechanism comprising of update (Zt) and reset gate (Rt) whereas eGRU is restricted for single gate i.e. only update (Zt), wherein both the gates are in hidden state (Ht). The complexity of GRU is even more reduced by utilizing Soft sign (S1) activation function in eGRU instead of Sigmoid (S2) and Tanh (S3) in GRU.

[0063] The hybrid architecture combines CNN's feature extraction skills and eGRU's sequence modelling capabilities. The CNN layers are good at extracting spatial features from input data because they use convolutional filters to detect local patterns. In contrast, eGRU layers excel at capturing temporal dependencies by modeling sequential data and learning long-term dependencies. The combination of CNN and eGRU provides multiple benefits in vehicle speed prediction. CNN layers can derive spatial information such as road textures, curves, and other relationships between front and rear wheel speeds, indicators from the input dataset. This enables the system to learn discriminative characteristics that are useful for speed estimation. eGRU layers, on the other hand, can capture temporal dynamics in vehicle speed data while compensating for the sequential character of the input. The hybrid system can capture local spatial patterns and temporal relationships by integrating CNN and eGRU. This hybrid strategy allows the system to understand complicated correlations between input data and target variables, which improves prediction accuracy.

[0064] The proposed hybrid architecture includes a plurality of CNNs followed by a plurality of eGRUs and one FC layer. The CNN layers extract features from the input data, and the eGRU layers capture temporal dependencies. CNN layers typically comprise of convolutional layers. The convolutional layers filter the input data and convolve it across spatial dimensions to extract local features. The output from the CNN layers is then passed into the eGRU layers, comprising of recurrent units that record sequential dependencies. The eGRU layers use gating methods to regulate the input flow, allowing the system to selectively update the data at each timestep.

[0065] The output from the eGRU layers passes through the FC layer so as to provide required output/result in the form of prediction regarding Vehicle Speed.

System Training and Evaluation:

[0066] In the system training, a plurality of CNN, a plurality of eGRU and one fully connected layer have been employed. The hybrid architecture is trained with an appropriate optimisation technique, such as Adam, and the system weights are updated using backpropagation through time (BPTT). The loss function used in training is determined by the prediction task's unique objectives, such as an MSE loss function adapted to the situation. During training, the system learns to minimise the difference between predicted speed and actual vehicle speed values. The performance of the hybrid Embedded GRU architecture is assessed using appropriate evaluation measures such as mean absolute error (MAE), mean squared error (MSE), RMSE, R-squared, and EVS. These metrics provide information on the accuracy and precision of the estimated vehicle speed system. Comparative analysis can be used to compare the hybrid architecture's performance to that of other techniques, such as 4L-LSTM, 4L-GRU, and 4L-eGRU.

[0067] According to the invention, the hybrid eGRU architecture has been deployed on the resource constrained platforms or edge devices using float32 bit and fixed32 bit architectures.

[0068] Now, reference may be made to figure 3 indicating flow chart regarding working of the present invention and figure 4 depicting exemplary embodiment of the figure 3 as discussed hereinabove.

RESULTS:

[0069] The system architectures is evaluated with MAE, R2 Loss, RMSE, and EV scores. The system got an MAE of 0.509, RMSE of 1.072, R2 loss of 0.998, and explained variance score of 0.998.
[0070] The proposed methodologies are deployed on AURIX TC277 and LPC 1768 MCUs. On AURIX TC277 MCU, the energy consumption was 0.204 mJ with a latency of 0.386 ms for float32 architectures; for fixed32 architecture, the latency was 0.757 ms with an energy consumption of 0.399mJ. On LPC1768 MCU, the energy consumption was 1.760 mJ with a latency of 3.322 ms for float32 architectures; for fixed32 architecture, the latency was 0.53 ms with an energy consumption of 0.28 MJ. These measures are best among all the other available DNN architectures like 4L-LSTM, 4L-GRU, and 4L-eGRU.

[0071] We have compared our methodology with the state-of-the-art RNN models like 4L-LSTM, 4L-GRU, and 4L-eGRU, and in comparison, our system is at least 5.44X energy, 20.98X latency on Aurix TC277 MCU, 11.64X energy, 11.80X latency on LPC1768 with an increased prediction accuracy (i.e., MAE has reduced from 0.65 to 0.50) for a single-precision format with the best methodologies available. The table below shows the error for the present invention proposed architecture:

System Architecture Loss Functions
MAE MSE RMSE R2 EVS
Hybrid eGRU 0.509 1.149 1.072 0.998 0.998

EXPERIMENT RESULTS:

[0072] The proposed architecture for vehicle speed prediction tasks gives 0.5 mean absolute error. Our methodology on hardware gives at least an 8x reduction in latency and a 7x reduction in energy consumption compared with the state-of-the-art methodology like 4L-LSTM, 4L-GRU, and 4L-eGRU as shown in literature 1.

ADVANTAGES OF THE INVENTION:

-Energy efficient and highly accurate for vehicle speed prediction with only four-wheel speed data on a resource-constrained environment, which causes an increase in driving range per mile, i.e., vehicle mileage;

- It involves simple architecture;

-It serves its function effectively.

APPLICATION OF THE INVENTION:-

-Autonomous vehicles,
- Electric vehicles,
-Intelligent driver assistance systems,
-Intelligent transportation systems.

[0073] It is best suited for vehicle speed prediction where the resources are the major concern from the hardware perspective without neglecting accuracy.

[0074] Each of the appended claims defines a separate invention, which for infringement purposes is recognized as including equivalents to the various elements or limitations specified in the claims. Depending on the context, all references below to the "invention" may in some cases refer to certain specific embodiments only. In other cases, it will be recognized that references to the "invention" will refer to subject matter recited in one or more, but not necessarily all, of the claims.

[0075] Groupings of alternative elements or embodiments of the invention disclosed herein are not to be construed as limitations. Each group member can be referred to and claimed individually or in any combination with other members of the group or other elements found herein. One or more members of a group can be included in, or deleted from, a group for reasons of convenience and/or patentability. When any such inclusion or deletion occurs, the specification is herein deemed to contain the group as modified thus fulfilling the written description of all groups used in the appended claims.

[0076] It will be understood by those within the art that, in general, terms used herein, and especially in the appended claims (e.g., bodies of the appended claims) are generally intended as "open" terms (e.g., the term "including" should be interpreted as "including but not limited to," the term "having" should be interpreted as "having at least," the term "includes" should be interpreted as "includes but is not limited to," etc.). It will be further understood by those within the art that if a specific number of an introduced claim recitation is intended, such an intent will be explicitly recited in the claim, and in the absence of such recitation no such intent is present. For example, as an aid to understanding, the following appended claims may contain usage of the introductory phrases "at least one" and "one or more" to introduce claim recitations. However, the use of such phrases should not be construed to imply that the introduction of a claim recitation by the indefinite articles "a" or "an" limits any particulars claim containing such introduced claim recitation to inventions containing only one such recitation, even when the same claim includes the introductory phrases "one or more" or "at least one" and indefinite articles such as "a" or "an" (e.g., "a" and/or "an" should typically be interpreted to mean "at least one" or "one or more"); the same holds true for the use of definite articles used to introduce claim recitations. In addition, even if a specific number of an introduced claim recitation is explicitly recited, those skilled in the art will recognize that such recitation should typically be interpreted to mean at least the recited number (e.g., the bare recitation of "two recitations," without other modifiers, typically means at least two recitations, or two or more recitations). Furthermore, in those instances where a convention analogues to "at least one of A, B and C, etc." is used, in general such a construction is intended in the sense one having skill in the art would understand the convention (e.g., "a system having at least one of A, B and C" would include but not be limited to systems that have A alone, B alone, C alone, A and B together, A and C together, B and C together, and/or A, B, and C together, etc.). In those instances where a convention analogous to "at least one of A, B, or C, etc." is used, in general such a construction is intended in the sense one having skill in the art would understand the convention (e.g., "a system having at least one of A, B, or C" would include but not be limited to systems that have A alone, B alone, C alone, A and B together, A and C together, B and C together, and/or A, B, and C together, etc.). It will be further understood by those within the art that virtually any disjunctive word and/or phrase presenting two or more alternative terms, whether in the description, claims, or drawings, should be understood to contemplate the possibilities of including one of the terms, either of the terms, or both terms. For example, the phrase "A or B" will be understood to include the possibilities of "A" or "B" or "A and B".

[0077] The above description does not provide specific details of manufacture or design of the various components. Those of skill in the art are familiar with such details, and unless departures from those techniques are set out, techniques, known, related art or later developed designs and materials should be employed. Those in the art are capable of choosing suitable manufacturing and design details.

[0078] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. It will be appreciated that several of the above-disclosed and other features and functions, or alternatives thereof, may be combined into other systems or applications. Various presently unforeseen or unanticipated alternatives, modifications, variations, or improvements therein may subsequently be made by those skilled in the art without departing from the scope of the present disclosure as encompassed by the following claims.

[0079] The claims, as originally presented and as they may be amended, encompass variations, alternatives, modifications, improvements, equivalents, and substantial equivalents of the embodiments and teachings disclosed herein, including those that are presently unforeseen or unappreciated, and that, for example, may arise from applicants/patentees and others.

[0080] While various aspects and embodiments have been disclosed herein, other aspects and embodiments will be apparent to those skilled in the art. The various aspects and embodiments disclosed herein are for purposes of illustration and are not intended to be limiting, with the true scope and spirit being indicated by the following claims.
, Claims:We Claim

1.A method for predicting a vehicle speed, the method comprising steps of:
- continuously monitoring speed of a plurality of wheels of a vehicle to generate a vehicle speed data (100A);
- processing the vehicle speed data (100A) of the plurality of wheels with breaking and non-breaking conditions on different types of roads through a data processor (101) for refining the data to structured data in a tensor format thereby forming a processed data;
- feeding the processed data to a plurality of convolution (CNN) layers sequentially for filtration to disregard redundant information from the processed data and to generate an output containing filtered data constituting containing key vehicle speed characteristics;
- feeding the output of a last layer of the plurality of convolution layers to a plurality of eGRU layers in a sequence for condensing the output data thereby generating a condensed data;
- providing the condensed data from a last layer of the plurality of eGRU layers to a connected dense layer (104), the connected dense layer transforms the condensed data having multi-dimensional information into a single and continuous value thereby predicting the vehicle speed.
2.The method as claimed in claim 1, wherein the processed data is fed to a first convolution (CNN) layer (102A) of the plurality of convolution layers to generate a filtered data as a first output, the first output is inputted to a subsequent convolution layer (102B) of the plurality of convolution layers for performing convolution to generate a further filtered data as a subsequent output;
wherein the subsequent output of the subsequent convolution layer (102B) is fed as input to a first eGRU layer (103A) for processing the data sequentially across time series to generate a first output of the first eGRU layer (103A) of the plurality of eGRU layers;
wherein the first output of the first eGRU layerl (103A) of the plurality of eGRU layers is applied as an input to a subsequent eGRU layer (103B) of the plurality of eGRU layers for refining temporal relationships between features and condensing the information for the prediction;
and subsequent output of the subsequent eGRU layer/cell (103B) is provided as input to the connected dense layer (104) to obtain the prediction speed of the vehicle (105).

3.The method as claimed in claim 1 or claim 2, wherein the convolution layer constitutes a one dimensional layer.

4. The method as claimed in any one of claims 1-3, wherein the vehicle speed data is monitored on a variety of road surfaces, including dry road, straight dry asphalt, ice, snow, and curve circles.

5. The method as claimed in any one of claims 1-4, wherein the eGRU layers , each of which comprises of an update gate (Zt) and has soft sign (S1) as their activation function.

6. The method as claimed in anyone of claims 1-5, wherein the step of processing the vehicle speed data includes identifying missing values, outliers, and discrepancies; and wherein the missing values are imputed using techniques comprising mean imputation and interpolation.

7. The method as claimed in any one of claims 1-6, wherein the vehicle speed data comprises of a speed data of a front left (FL), a front right (FR), a rear left (RL), and a rear right (RR) wheels as input.

8. The method as claimed in any one of claims 1-7, wherein the plurality of convolution (CNN) layers derive a spatial information including road textures, curves, and other relationships between the speed of the front and the rear wheels, in which the plurality of eGRU layers capture temporal dynamics in vehicle speed data while compensating for the sequential character of the input.

9. The method as claimed in claim 8, wherein the plurality of convolution (CNN) layers filter the input data and convolve it across spatial dimensions to extract local features, in which the output from the last layer of the plurality of CNN layers is passed into the eGRU layers, comprising of recurrent units that record sequential dependencies.

10. The method as claimed in claim 9, wherein the eGRU layers use gating methods to regulate the input flow, allowing the system to selectively update the data at each timestep.

11. A system for predicting a vehicle speed (100) comprising of data processor (101) for refining data to structured data in tensor format thereby forming a processed data in communication with hybrid Embedded GRU architecture, wherein the hybrid Embedded GRU architecture includes a plurality of convolution layers (102A,102B) for filtration of the processed data to disregard redundant information from the processed data and to generate an output containing filtered data constituting key vehicle speed characteristics, a plurality of embedded GRU layers (103A,103B) for condensing filtered data and at least one dense layer (104) for transforming the condensed data having multi-dimensional information into a single and continuous value thereby predicting the vehicle speed, in which the data processor (101) is fed with at least four wheel vehicle speed data (101A) with breaking and non-breaking conditions on different types of roads.

12.The system as claimed in claim 11, wherein the convolution layer constitutes one dimensional layer.

13. The system as claimed in claim 11 or 12, wherein the vehicle speed data is monitored on a variety of road surfaces, including dry road, straight dry asphalt, ice, snow, and curve circles.

14. The system as claimed in claims 11-13, wherein the eGRU layers , each of which comprises of an update gate (Zt) and has soft sign (S1) as their activation function.

15. The system as claimed in claims 11-14, wherein the vehicle speed data includes identifying missing values, outliers, and discrepancies; and wherein the missing values are imputed using techniques comprising mean imputation and interpolation.

16. The system as claimed in claims 11-15,wherein the vehicle speed data comprises of a speed data of a front left (FL), a front right (FR), a rear left (RL), and a rear right (RR) wheels as input.

17. The system as claimed in claims 11-16, wherein the plurality of convolution (CNN) layers derive a spatial information including road textures, curves, and other relationships between the speed of the front and the rear wheels.

18. The system as claimed in claim 17, wherein the plurality of convolution (CNN) layers filter the input data and convolve it across spatial dimensions to extract local features, in which the output from the CNN layers is passed into the eGRU layers, comprising of recurrent units that record sequential dependencies.






19. The system as claimed in claim 18, wherein the plurality of eGRU layers use gating methods to regulate the input flow, allowing the system to selectively update the data at each timestep.

Documents